AI automation cost depends almost entirely on what you are automating, and barely at all on the AI. The same request, automate our invoicing, can be a small script or a six-figure project.
Ask three vendors to automate one process and the quotes can disagree by an order of magnitude. They are usually pricing different amounts of work you cannot see yet.
You are trying to budget for something with no sticker price, and most guides hand you a number that falls apart the moment it meets your real systems. This article skips the fake price list. It explains what genuinely moves the cost, so you can read a quote and tell real work from padding.
- The model is rarely the expensive part. Connecting it to your tools, cleaning up messy inputs, and handling the cases where it gets things wrong are where the money goes.
- The same automation can be cheap or costly depending on how many systems it touches and how much a wrong output costs you.
- No-code starts cheaper and hits a ceiling. Custom costs more upfront and removes the ceiling. Both carry a maintenance cost people forget.
- Most of the price is set before any code is written, at the scoping stage. That is also where you have the most control over it.
- You lower the cost by narrowing the scope, rather than by negotiating the rate.
The AI is the smallest line on the invoice
People expect the model to be the expensive part. Model API calls are cheap, often cents per run, and they keep getting cheaper. The cost sits in everything wrapped around the model.
Wiring the automation into your CRM, your inbox, or your accounting tool takes engineering. Teaching it to handle the input that does not fit the happy path takes more. Deciding when a human signs off before anything goes out takes more again. That work is roughly the same whether the model behind it is large or small.
This is why webvise prices AI workflow automation per workflow after a short discovery call, rather than quoting the AI by the hour. A first custom workflow usually lands in the low five figures to build, and the actual ranges plus the payback math behind them sit in how to calculate AI automation ROI.
Why the same request gets very different quotes
Four things explain most of the gap between a cheap automation and an expensive one. The AI is not one of them.
- How many systems it touches. A one-way webhook into a single app is quick. Two-way sync across a CRM, an ERP, and a database is a different size of job.
- How clean the inputs are. Structured data the automation can trust keeps the build short. Scanned PDFs, free-text emails, and inconsistent naming add a cleanup layer before the real work starts.
- What a wrong output costs. An automation that drafts something for a human to check is cheap to get wrong. One that writes to your finance system or sends to a client needs review steps, approvals, and an audit trail, and those guardrails are most of the build.
- How often it runs. Constant high volume costs more to run each month than an occasional job, and it often forces a larger model. The model choice is a budget decision, made when the work is scoped.
The same drivers, read as a quick gut check:
| Driver | Cheaper end | Pricier end |
|---|---|---|
| Systems | One app, one direction | Several systems, two-way sync |
| Inputs | Clean, structured data | Scanned or free-text, needs cleanup |
| Stakes | A human reviews the output | Writes to production, needs guardrails |
| Volume | Occasional runs | Constant, high volume |
| Ownership | Vendor maintains it | Brittle no-code your team babysits |
A documentary producer who pitches commissions to broadcasters shows the stakes effect. The automation webvise built generates a broadcaster-ready document from an idea, in his own voice, with every claim traced to a source. It cost more than a plain data sync would have, because one off-tone page can lose a commission, so the output had to be exact every time. The model was the easy part; the voice fidelity and the exact formatting were the real work.
No-code or custom changes the shape of the cost
No-code tools like Zapier, Make, and n8n are cheaper to start and fast to ship. They carry monthly platform fees, a ceiling on logic and volume, and the hours your team spends babysitting flows. A team running several no-code subscriptions often still loses an afternoon a month replaying failed runs across two dashboards.
A custom build costs more upfront and takes longer. In return it owns its own maintenance, has no ceiling for the work it was scoped for, and stops billing you in your team's evenings. Which one wins depends on how long the workflow will live and how much it will grow. The build versus buy decision and the no-code comparison go through that choice in detail.
Most of the cost is decided before any code is written
By the time someone quotes you, the expensive decisions are already made. The biggest one is whether the workflow is worth automating at all. Pressure-test it against four questions before you pay anyone.
- Volume. Does it run often enough, dozens of times a month, for the build to pay back? A rare task rarely earns its cost.
- Variance. Is the work rule-shaped or judgment-shaped? Clear rules are cheap to automate. Constant judgment calls are not.
- Stability. Will the systems it depends on still be here in a year? Automating a process you are about to replace burns the build.
- Ownership. Who fixes it when it breaks? An automation with no owner becomes an expensive thing nobody trusts.
Scoping is the cheapest place to cut cost, and it is the first thing webvise does. A short discovery call turns a vague request into a fixed build price, so you commit to a number instead of a range. The task that wastes the most hours every week is almost always the right one to start with.
AI automation cost only becomes a real number once the work is scoped. To get yours, tell webvise the task that eats the most time through the contact form, and you will get a fixed price for it instead of a guess.
Development practices are aligned with ISO 27001 and ISO 42001 standards.